From c6124aa46f5d2abcfd55dae9c4d1f9103b402561 Mon Sep 17 00:00:00 2001
From: Lumi AI Workspace <hong@cloudjetkpi.com>
Date: Tue, 7 Jul 2026 00:31:22 +0700
Subject: [PATCH] docs: add benchmark slice register

---
 docs/evidence/BENCHMARK_SLICE_REGISTER.md | 92 +++++++++++++++++++++++
 1 file changed, 92 insertions(+)
 create mode 100644 docs/evidence/BENCHMARK_SLICE_REGISTER.md

diff --git a/docs/evidence/BENCHMARK_SLICE_REGISTER.md b/docs/evidence/BENCHMARK_SLICE_REGISTER.md
new file mode 100644
index 00000000..80beb296
--- /dev/null
+++ b/docs/evidence/BENCHMARK_SLICE_REGISTER.md
@@ -0,0 +1,92 @@
+# Benchmark Slice Register
+
+OpenAMP Foundry should track benchmark slices explicitly.
+
+Aggregate benchmark performance is rarely enough. A component can look strong overall while failing on the slice that matters most for safe selection, novelty, feasibility, or review readiness.
+
+This register defines how to name and preserve benchmark slices without adding biological protocols, candidate-generation instructions, or optimization objectives.
+
+## Why this matters
+
+Benchmarks become misleading when failure slices are unnamed.
+
+If the repo only remembers aggregate results, future contributors cannot tell whether a change improves robust selection or merely shifts performance across hidden subgroups.
+
+Named slices make benchmark claims harder to launder through averages.
+
+## What is a benchmark slice?
+
+A benchmark slice is a declared subset of an evaluation used to inspect behavior under a meaningful boundary.
+
+Examples of slice dimensions:
+
+- source database or curation path;
+- dataset version;
+- novelty band;
+- feasibility band;
+- safety-review band;
+- sequence length band;
+- known-control group;
+- external-predictor availability;
+- review-readiness level;
+- time or reference-version boundary.
+
+A slice should exist because it tests a failure mode, not because it makes a result look better.
+
+## Minimum slice record
+
+```markdown
+## Benchmark slice
+
+**Slice ID:** Stable identifier.
+
+**Purpose:** What failure mode or claim boundary this slice tests.
+
+**Inclusion rule:** How records enter the slice.
+
+**Exclusion rule:** How records are excluded.
+
+**Dataset or reference version:** Version, commit, date, or artifact.
+
+**Primary metric:** Metric used for this slice.
+
+**Expected behavior:** What should a useful component do on this slice?
+
+**Known caveat:** Why this slice could mislead.
+
+**Owner:** Maintainer, reviewer, or workstream responsible for updates.
+```
+
+## Reviewer guidance
+
+Reviewers should ask for a slice record when aggregate performance hides a meaningful boundary.
+
+A slice does not need to be permanent. It should remain while it captures a live failure mode or claim boundary.
+
+## Agent guidance
+
+Agents should not invent slices after seeing which ones look favorable.
+
+An agent proposing a new slice should state the failure mode first, then define the inclusion rule.
+
+## Anti-patterns
+
+Avoid:
+
+- creating slices only after observing results;
+- reporting only the best slice;
+- changing slice definitions without recording why;
+- comparing components across different slice definitions;
+- using tiny slices without marking uncertainty;
+- hiding weak slices behind aggregate improvement.
+
+## Acceptance test
+
+A benchmark slice is useful when a reviewer can answer:
+
+1. What failure mode does this slice test?
+2. How are records included and excluded?
+3. Which claim boundary depends on it?
+4. What caveat travels with the result?
+
+If those answers are missing, the slice is not benchmark governance. It is decoration.